Editing an AI avatar's nose and keeping the same person: Nano Banana Pro vs Google Flow
Updated 30 October 2026 · 5 min read
These are the findings from our own tests while building ShapesShifter — what didn't work, the one change that made it work, and how far each generator actually moves a feature. We changed one variable at a time.
Finding 1: a reference shape alone often changes nothing
We drew three different nose shapes over the same avatar, at roughly the size of the original nose, and used the same short prompt for all three. The results came back practically identical to the original. The generator saw a shape close to what was already there and kept the face.
Finding 2: one sentence of description makes it work
Same three images, same prompt, with one sentence added describing the character of each shape (for example "a straight, narrow bridge with a softly rounded tip"). The three results became clearly different from each other and from the original, while the rest of the face held. This is why every reference shape in ShapesShifter comes with its own written description that is pasted into the prompt automatically.
Finding 3: longer prompts break identity
Very long prompts (around 8,000 characters, with many rules) diluted the "keep the same person" instruction and the generator began changing the person instead of the nose. Prompts of roughly 900 characters kept identity. Rule we kept: add an instruction only when you can see the defect it fixes.
Finding 4: blue outline vs skin-coloured overlay
| Reference style | Making it bigger | Making it smaller | Consistency |
|---|---|---|---|
| Blue outline | Works | Works — the only style that did | Consistent |
| Skin-coloured overlay | Sometimes a larger change | Fails | Two identical runs gave different results |
When shrinking, a skin-coloured shape doesn't show where the new outline ends and the old feature begins. Blue makes that boundary obvious.
Finding 5: how far each generator goes
| Generator | Behaviour with a reference | Approximate range (nose width) |
|---|---|---|
| Nano Banana Pro (direct) | Reads the reference as a direction, applies a standard change and stops | ≈ 10–13%, whatever you ask |
| Google Flow, agent on | Scales with the reference | up to ≈ +45% wider, ≈ −40% narrower; stops following past ≈ +70% |
So for small refinements either works; for a clearly different nose, use Flow with the agent turned on.
Finding 6: asymmetric shapes need retries
A deviated nose or uneven lips usually took 2–3 generations. Image models are trained mostly on symmetric faces and pull results back toward symmetry. That's a property of the generator, not of the reference.
Practical recipe
- Front-facing, evenly lit source image.
- One feature per edit, blue outline of the target shape on the image.
- Short prompt: change only this feature + one sentence describing the shape + keep everything else identical.
- Google Flow with the agent on for bigger changes; expect 2–3 tries on asymmetric shapes.
- Continue the next edit from the best result, never from the original.
The full step-by-step method is in How to change one facial feature without changing the face.
112 reference shapes, each with its description, exported with a ready prompt.
Try ShapesShifterFAQ
Is Nano Banana Pro or Google Flow better for changing one facial feature?
In our tests Google Flow with the agent on followed the reference much more closely. Nano Banana Pro alone applied about 10–13% change whatever was asked.
What colour should the reference outline be?
Blue worked in both directions. Skin-coloured overlays were inconsistent and failed when shrinking a feature.
Why did the AI ignore my reference shape?
If the shape is close to the original and the prompt only says "follow the reference", results come back nearly unchanged. Add one sentence describing the target shape.